Researchers at Monash University’s Pharmaceutical Sciences department have used machine learning algorithms to identify several promising antibiotic candidates effective against drug-resistant bacteria, reducing the time for initial screening from years to months.

The team trained neural networks on databases of known antibiotics and their molecular properties, then used the trained models to predict which molecules from libraries containing billions of compounds might have antibiotic activity. Laboratory testing confirmed that 12 of the top 20 predictions showed antibacterial effects.

Antibiotic resistance represents one of the most serious public health threats globally. Common bacterial infections increasingly resist available antibiotics, and pharmaceutical companies have largely abandoned antibiotic development because it’s not commercially attractive. New approaches to antibiotic discovery are desperately needed.

The Discovery Challenge

Developing new antibiotics traditionally involves screening thousands of chemical compounds against bacteria to find ones that kill or inhibit growth. This process is slow, expensive, and often fails to find novel antibiotic classes because screening focuses on chemical structures similar to existing antibiotics.

Nature provided most antibiotics currently in use, discovered by screening microorganisms for compounds they produce to compete with other bacteria. But easy discoveries are exhausted, and finding new natural antibiotics requires sampling ever more diverse environments with decreasing success rates.

Chemical synthesis can create compounds that don’t exist in nature, but the number of possible molecular structures is effectively infinite. Screening even a tiny fraction exceeds available resources. Computational prediction offers a way to focus experimental work on molecules likely to have desired properties.

Professor Jian Li, who leads the Monash research group, said machine learning excels at pattern recognition tasks where rules are complex and partially understood. Antibiotic activity depends on molecular shape, charge distribution, and interactions with bacterial cellular machinery in ways that are difficult to predict from first principles.

The Computational Approach

The team used public databases containing information on thousands of known antibiotics and their molecular structures. They also incorporated data on compounds tested for antibiotic activity that proved inactive. This negative data helps models learn what doesn’t work, which is almost as valuable as knowing what does.

The neural networks learned to associate molecular features like functional groups, overall shape, and electronic properties with antibiotic activity. Once trained, the models evaluated new molecular structures and predicted the probability each would show antibacterial effects.

The researchers then scored billions of molecules from commercially available chemical libraries using the trained models. They prioritised compounds predicted to be active but structurally dissimilar to existing antibiotics, since novel structures are more likely to work through different mechanisms and avoid existing resistance.

Purchasing and testing 100 top-ranked compounds cost about $50,000 in materials and lab time, vastly cheaper than screening that many compounds randomly. Twelve compounds showed antibacterial activity, a hit rate far exceeding traditional screening.

Further testing characterised those hits, determining which bacteria they affected, at what concentrations, and through what mechanisms. Several compounds showed activity against antibiotic-resistant strains, including MRSA and multi-drug-resistant Acinetobacter.

From Hits to Drugs

Finding compounds with antibacterial activity is just the beginning of drug development. Candidate antibiotics must be non-toxic to human cells, remain stable in the body long enough to be effective, penetrate infected tissues, and not be rapidly broken down by metabolism.

Most compounds with promising antibacterial activity fail at subsequent development stages. They might be toxic, ineffective when tested in animals, or unstable when formulated as medicines. Developing a single approved antibiotic typically requires testing thousands of compounds and costs hundreds of millions of dollars.

The Monash compounds are very early-stage candidates. They’ve shown antibacterial activity in laboratory tests, but animal testing, toxicity studies, and chemical optimisation remain ahead. The research group is conducting additional experiments to understand how the compounds work and which show the most promise.

Some of the active compounds have chemical structures that suggest modifications could improve their properties. Medicinal chemists can synthesise variants of the original structures, testing whether changes enhance antibacterial activity or improve drug-like properties.

This iterative optimisation is another area where machine learning might help. Rather than chemists synthesising and testing variants based on intuition and experience, algorithms can suggest modifications predicted to improve specific properties. The Monash team is exploring this direction.

Pharmaceutical Industry Context

Major pharmaceutical companies mostly abandoned antibiotic research decades ago because the economics don’t work. Antibiotics are used for short periods to treat infections, unlike chronic disease medications taken daily for years. Resistance development limits the useful lifespan of new antibiotics. And pricing pressure from governments and health insurers limits revenues.

Most current antibiotic development happens in small biotech companies or academic labs, funded by government grants and non-profit organisations. This model produces scientific insights but rarely leads to approved drugs because later-stage development and clinical trials are expensive.

Several initiatives aim to improve antibiotic development economics. Prize funds for successful new antibiotics, changes to regulatory requirements, and guaranteed purchase commitments from governments all attempt to make development more attractive. Results have been mixed.

Academic research like the Monash work typically identifies early-stage candidates that are then licensed to companies for further development. Universities lack the resources and expertise for clinical trials and manufacturing. Whether industry partners emerge depends on how promising the compounds appear after additional testing.

Machine Learning in Drug Discovery

Pharmaceutical companies and research groups worldwide are increasingly applying machine learning to drug discovery. Applications range from target identification to molecule generation to predicting clinical trial outcomes.

Results have been promising in some areas and disappointing in others. Machine learning excels at analysing large datasets to find patterns, but drug development involves small datasets relative to the complexity of biological systems. Models trained on limited data often fail to generalise.

There’s also the question of what problem machine learning solves. If the bottleneck in drug development is identifying initial candidates, computational approaches help. But if bottlenecks are toxicity, efficacy in humans, or development costs, computational methods have less impact.

The Monash work demonstrates that machine learning can identify antibiotic candidates more efficiently than random screening. Whether those candidates progress to useful drugs depends on subsequent development stages where computational approaches help less.

For organisations trying to evaluate when AI-driven drug discovery will impact their sector, understanding both capabilities and limitations of current approaches matters. The technology accelerates some aspects of discovery but doesn’t eliminate the long, uncertain, expensive path from initial candidate to approved medicine.

Global Antibiotic Resistance

While the Monash research represents scientific progress, the antibiotic resistance problem continues worsening. Bacteria evolve resistance faster than new antibiotics are developed, and inappropriate antibiotic use accelerates resistance emergence.

Addressing antibiotic resistance requires both new drugs and better stewardship of existing antibiotics. Reducing unnecessary prescriptions, improving infection control in hospitals, and minimising agricultural antibiotic use all matter as much as discovering new drugs.

Some experts argue that developing narrow-spectrum antibiotics targeting specific pathogens makes more sense than broad-spectrum drugs that kill many bacterial species. Narrow-spectrum antibiotics reduce collateral damage to beneficial bacteria and potentially slow resistance development. But they require rapid, accurate diagnosis to identify which pathogen is causing infection.

Others advocate for renewed focus on preventing bacterial infections through vaccines and improved hygiene rather than treating them with antibiotics. Prevention is always preferable to treatment, but infections will continue occurring and antibiotics will remain necessary.

The Monash research contributes to the discovery side of the problem, demonstrating that machine learning can accelerate finding new antibiotic candidates. But it’s one piece of a much larger challenge that demands action on multiple fronts.